fix: use compaction as primary baseline, fix threshold, update marketing

[BYK]

Summary

Switches the eval comparison from tail-window to compaction (what real tools actually do), fixes the compaction threshold, and updates marketing copy with honest numbers from a fresh head-to-head run.

Eval changes

Remove tail-window from default baselines (run.ts)

Nobody uses raw tail-window in production — Claude Code, Codex, OpenCode all use compaction. Compaction is the honest baseline.

Fix compaction threshold (baselines.ts)

The first compaction was triggering at 80K total tokens. Real tools don't compact until ~140K (83.5% of the 200K context window minus output reserve). Fixed: compaction now only triggers when total > compactionThreshold (~140K), not at the tail budget (80K).

Fresh eval results (CM-1, 400K inflation, head-to-head)

Difficulty	Lore	Compaction	Delta
Easy	4.7/5	4.8/5	-2%
Medium	4.8/5	4.0/5	+19%
Hard	4.9/5	4.7/5	+5%
Average	4.8/5	4.5/5	+7%
Perfect scores	12/15	9/15	—

Lore's advantage is on mid-session details (decision alternatives, exact error messages, rejected approaches) that compaction summarizes away through repeated compression cycles.

Marketing updates

README.md

• Context retention table: compaction as primary comparison (not tail-window)
• Preference recall table: noted baselines pending re-run with compaction
• Summary: 4.8/5 with 12/15 perfect scores

docs/index.html

• Hero stat: "12/15 Perfect Scores at 400K Tokens"
• Detail retention: 4.8/5

Version history

• v6: updated to reference compaction comparison

Tests

• 1752 pass, 0 fail
• Typecheck clean